Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A system comprising: a processor and a memory of a computing device; and at least one module on the computing device configured to cause the processor to: receive an output of a face detector, the face detector detecting faces using Haar features, the output comprising a set of regions that the face detector has determined include faces; extract from each region of the set of regions, at least one descriptor of a not-Haar feature; and send the at least one descriptor of the not-Haar feature to a post-filter classifier placed at a terminating end of a face detection pipeline to categorize the output as a face or false positive, wherein the post-filter classifier is a trained support vector machine, wherein the post-filter classifier is trained using a set of training regions output from the face detector, the set of training regions including regions that do include a face and regions that are mistakenly classified by the face detector as including faces.
A face detection system uses a processor and memory to identify faces in images. It receives region proposals from a face detector (like Viola-Jones) that initially finds faces using Haar features. For each proposed region, it extracts a "not-Haar" feature descriptor (e.g., Local Binary Patterns, or facial recognition features). This descriptor is fed to a post-filter classifier, specifically a trained Support Vector Machine (SVM), positioned at the end of the face detection pipeline. The SVM was pre-trained on a dataset of regions output by the initial face detector, labeled as either "face" or "not a face" (false positive), allowing it to reduce false positives.
2. The system of claim 1 , wherein the face detector is a Viola Jones face detector.
The face detection system described previously, which uses a post-filter classifier (SVM) to reduce false positives from an initial face detector that uses Haar features, specifically uses a Viola-Jones face detector to provide the initial region proposals. The system extracts non-Haar features from these regions before feeding them to the SVM post-filter.
3. The system of claim 1 , wherein the not-Haar feature is a Local Binary Pattern feature.
The face detection system described previously, which uses a post-filter classifier (SVM) to reduce false positives from an initial face detector that uses Haar features, uses Local Binary Pattern (LBP) features as the "not-Haar" feature extracted from the region proposals. This LBP feature is input to the SVM post-filter to determine if a region truly contains a face.
4. The system of claim 1 , where the not-Haar feature is a facial recognition feature.
The face detection system described previously, which uses a post-filter classifier (SVM) to reduce false positives from an initial face detector that uses Haar features, uses a facial recognition feature as the "not-Haar" feature extracted from the region proposals. This facial recognition feature is input to the SVM post-filter to determine if a region truly contains a face.
5. The system of claim 1 , wherein each region of the set of regions has an associated score wherein the associated score comprises an indication of a degree of confidence that a region associated with the score comprises a face.
The face detection system described previously, which uses a post-filter classifier (SVM) to reduce false positives, also uses a confidence score generated by the initial face detector for each region. This score represents the detector's certainty that the region contains a face. This score can be used by the SVM or another part of the pipeline.
6. A method comprising: receiving a region by a processor of a computing device, the region comprising an output of a face detector, the face detector detecting faces using Haar features; extracting a descriptor of a not-Haar feature from the received region; receiving the extracted descriptor; and determining results comprising a result indicating that the received region comprises a face or a result indicating that the received region does not comprise a face by sending the descriptor of the not-Haar feature to a post-filter classifier placed at a terminating end of a face detection pipeline, wherein the post-filter classifier is a trained support vector machine, wherein the post-filter classifier is trained using a training set of regions output from the face detector, the training set of regions including regions that do include a face and regions that are mistakenly classified by the face detector as including faces.
A method for face detection involves receiving a region of an image from a face detector (using Haar features). A "not-Haar" feature descriptor (e.g., Local Binary Patterns, or facial recognition features) is extracted from this region. The descriptor is then fed to a post-filter classifier, which is a trained Support Vector Machine (SVM) located at the end of the face detection pipeline. The SVM was trained on regions from the initial face detector, labeled as either containing a face or being a false positive, to determine whether the region actually contains a face or is a false positive.
7. The method of claim 6 , wherein the not-Haar feature is a facial recognition feature.
In the face detection method using a post-filter classifier (SVM) to reduce false positives, a facial recognition feature is used as the "not-Haar" feature extracted from the region, which comes from an initial face detector that uses Haar features. The facial recognition feature is then used to classify the region as a face or not.
8. The method of claim 6 , further comprising: extracting the descriptor of the face detection feature using a Local Binary Pattern feature extractor.
In the face detection method using a post-filter classifier (SVM) to reduce false positives, a Local Binary Pattern (LBP) feature extractor is used to extract the "not-Haar" feature descriptor from the region, which comes from an initial face detector that uses Haar features. The LBP descriptor is then used to classify the region as a face or not.
9. The method of claim 6 , wherein the region has an associated score wherein the associated score comprises an indication of a degree of confidence that the region comprises a face.
In the face detection method using a post-filter classifier (SVM) to reduce false positives, each region from the initial face detector has a score representing the face detector's confidence that the region contains a face. This score may be used in further processing or by the post-filter.
10. The method of claim 6 , further comprising: reducing a set of regions input to the post-filter classifier by a number of regions that the face detector has determined include faces.
The face detection method using a post-filter classifier (SVM) reduces the number of regions passed to the SVM. This is done by first using the initial face detector that uses Haar features, and only passing a subset of the regions it identifies to the post-filter, potentially skipping regions with high confidence scores.
11. The method of claim 6 , wherein the face detector is a Viola-Jones face detector.
In the face detection method using a post-filter classifier (SVM) to reduce false positives, the initial face detector, which provides regions of interest, is a Viola-Jones face detector. It uses Haar features and the SVM uses "not-Haar" features.
12. A computer-readable storage device comprising computer-executable instructions which when executed cause at least one processor to: use a not-Haar feature to detect a face in an image by: receiving the output of a face detector, the face detector detecting faces using Haar features, the output comprising a region; extracting from the region, a descriptor of a not-Haar feature; and sending the descriptor to a trained face detection post-filter classifier, wherein the trained face detection classifier provides a result indicating that the region includes a face or provides a result indicating that the region does not include a face, wherein the trained face detection post-filter classifier is a trained support vector machine, wherein the trained face detection post-filter classifier is trained using a set of training regions output from the face detector, the set of training regions including regions that do include a face and regions that are mistakenly classified by the face detector as including faces.
A computer-readable storage device holds instructions for face detection. These instructions cause a processor to receive region proposals from a face detector that uses Haar features. The processor extracts a "not-Haar" feature descriptor from each region. The descriptor is sent to a trained Support Vector Machine (SVM) post-filter classifier. This SVM outputs a result indicating whether the region contains a face or not. The SVM was trained on a dataset of regions from the initial face detector, labeled as either "face" or "not a face", to improve accuracy.
13. The computer-readable storage device of claim 12 , comprising further computer-executable instructions, which when executed cause the at least one processor to: receive the output of a Viola Jones face detector, the output comprising a region that the Viola Jones face detector determined includes a face.
The computer-readable storage device containing instructions for face detection, as described previously, receives region proposals from a Viola-Jones face detector, which is used as the initial face detector. The system extracts "not-Haar" features from these regions and filters them using an SVM.
14. The computer-readable storage device of claim 12 , comprising further computer-executable instructions, which when executed cause the at least one processor to: extract from the region, a descriptor of a not-Haar feature comprising a facial recognition feature.
The computer-readable storage device containing instructions for face detection, as described previously, uses a facial recognition feature as the "not-Haar" feature that is extracted from the region. This feature is then used as input into an SVM that acts as a post-filter.
15. The computer-readable storage device of claim 12 , comprising further computer-executable instructions, which when executed cause the at least one processor to: train the trained face detection post-filter classifier using Support Vector Machine training techniques.
The computer-readable storage device containing instructions for face detection, as described previously, trains the Support Vector Machine (SVM) post-filter classifier using standard Support Vector Machine (SVM) training techniques using labeled data. The SVM is trained on regions from the initial Haar-feature based face detector, labeled as either "face" or "not a face."
16. The computer-readable storage device of claim 12 , comprising further computer-executable instructions, which when executed cause the at least one processor to: execute the trained face detection post-filter classifier at a terminal end of a face detection pipeline.
The computer-readable storage device containing instructions for face detection, as described previously, executes the trained Support Vector Machine (SVM) post-filter classifier at the very end of the entire face detection pipeline, after the initial face detector (using Haar features) and feature extraction steps.
17. The computer-readable storage device of claim 12 , comprising further computer-executable instructions, which when executed cause the at least one processor to: receive the output of a Viola Jones face detector.
The computer-readable storage device containing instructions for face detection, as described previously, receives the output of a Viola-Jones face detector as the initial set of region proposals. The Viola-Jones face detector uses Haar features. These regions are then processed to extract "not-Haar" features before post-filtering with an SVM.
18. The system of claim 1 , wherein each region is labeled one of positive or negative, wherein a region labeled positive is a true positive, and wherein a region labeled negative is a false positive.
The face detection system, which uses a post-filter classifier (SVM) to reduce false positives, labels each region outputted from the initial face detector (which uses Haar features) as either "positive" (true positive, contains a face) or "negative" (false positive, does not contain a face). This labeling is used for training the SVM.
19. The method of claim 6 , wherein each region is labeled one of positive or negative, wherein a region labeled positive is a true positive and wherein a region labeled negative is a false positive.
The face detection method, which uses a post-filter classifier (SVM) to reduce false positives, labels each region from the initial face detector as either "positive" (true positive, contains a face) or "negative" (false positive, does not contain a face). This labeling is used for training the SVM. The initial face detector uses Haar features.
20. The computer-readable storage device of claim 12 , wherein each region is labeled either positive or negative, wherein a region is labeled positive is a true positive, and wherein a region labeled negative is a false positive.
The computer-readable storage device with instructions for face detection, as described previously, labels each region output from the initial face detector (which uses Haar features) as either "positive" (true positive, contains a face) or "negative" (false positive, does not contain a face). This labeled data is used to train the SVM.
Unknown
October 28, 2014
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.